Hugging Face Launches Microduck for Open-Source Robotics

Hugging Face is taking its open-source model philosophy into the physical world with Microduck, a $399 research robot built to let developers train and modify embodied AI systems. The company unveiled the duck-like machine on August 27, according to TechCrunch.

An affordable platform for physical AI

Microduck is a 25-centimeter-tall robot that can waddle, pick up objects weighing up to 800 grams, recover after falling, crouch, and roller skate. Its hardware combines a camera, lidar sensors, and two inertial measurement units, giving developers a compact platform for experimenting with perception, movement, and control.

The price matters because physical AI has often been locked behind expensive research hardware or proprietary software. Microduck is positioned as an accessible starting point: it ships before Christmas and is small enough for a lab, classroom, or serious hobbyist workspace.

From simulation to a trainable machine

The most important part of the launch is not the novelty of the form factor, but the development loop around it. Pollen Robotics says Microduck behaviors can be trained in simulation and deployed directly to the hardware. Developers can then fine-tune the system, retrain it, and redeploy it. The SDK, simulation environment, and full reinforcement-learning stack are available on GitHub, creating a path from an experiment to a repeatable physical behavior.

That workflow mirrors what software teams already expect from modern AI: versioned environments, measurable training runs, and rapid iteration. For builders, it creates a useful testbed for connecting models to sensors and actuators without starting with a large industrial platform.

Why builders should pay attention

Microduck also makes the open-source trade-off tangible. Open models can improve auditability and local control, but applications layered on top of them may still access cameras and microphones or send data to external services. Privacy therefore depends on the entire software stack, not just the model weights.

For n8n Lab readers, the broader signal is clear: agentic systems are moving from browser and API workflows toward environments where perception, feedback, and safe recovery are first-class concerns. The teams that learn to keep those loops observable and governed will have an advantage as physical AI becomes easier to prototype.

Source: TechCrunch, by Rebecca Bellan.